Amazon’s Graviton and Trainium scale up
Reports say Amazon’s Graviton CPUs and Trainium accelerators are rapidly scaling and now competing more directly with Intel, AMD and Nvidia in AI and infrastructure workloads. The coverage notes hyperscaler momentum behind custom silicon as these chips expand their performance footprint. (x.com)
Amazon’s in-house chips have moved from side project to core business: Chief Executive Andy Jassy said Amazon’s chips unit now runs at more than $20 billion a year. (aboutamazon.sg) A central processing unit is the general-purpose engine that runs databases, web servers, and business software; an artificial intelligence accelerator is a specialist chip built to train and serve large models. Amazon sells both through Amazon Web Services: Graviton for cloud computing and Trainium for artificial intelligence. (aws.amazon.com 1) (aws.amazon.com 2) Trainium 2 reached general availability on December 3, 2024, in Amazon Elastic Compute Cloud Trn2 instances. Each Trn2 instance uses 16 Trainium 2 chips and delivers up to 20.8 petaflops of FP8 compute, 1.5 terabytes of high-bandwidth memory, and 3.2 terabits per second of networking. (aws.amazon.com) Amazon also built a larger package for the same chip: Trn2 UltraServers connect 64 Trainium 2 chips across four instances with Amazon’s NeuronLink fabric. Amazon says that setup reaches up to 83.2 petaflops of FP8 compute and 6 terabytes of memory for training and inference on models with hundreds of billions to more than 1 trillion parameters. (aws.amazon.com 1) (aws.amazon.com 2) On the central processing unit side, Graviton has spread far beyond a niche Arm experiment. Amazon Web Services says more than 90,000 customers use Graviton, and Graviton-based instances cost up to 20% less than comparable x86 instances while using up to 60% less energy for the same performance. (aws.amazon.com) Graviton 4 became generally available in July 2024, and Amazon says the chip delivers four times the performance of the first Graviton generation. In December 2025, Amazon introduced Graviton 5 and said, for the third straight year, more than half of new central processing unit capacity added to Amazon Web Services was Graviton-based. (aboutamazon.com 1) (aboutamazon.com 2) Amazon is using the same playbook in artificial intelligence. Jassy said Amazon’s chips business, which includes Graviton, Trainium, and Nitro networking cards, is growing at triple-digit year-over-year rates, and Amazon said at re:Invent 2025 that Trainium 3 UltraServers were already becoming available. (aboutamazon.sg) (aboutamazon.com) The competitive target is clear even when Amazon does not name rivals in every sentence. Graviton replaces some workloads that once defaulted to Intel and Advanced Micro Devices processors, while Trainium is priced against Amazon’s own graphics-processing-unit instances: Amazon says Trn2 offers 30% to 40% better price-performance than its graphics-processing-unit-based P5e and P5en systems. (aws.amazon.com) (aws.amazon.com) Amazon’s pitch is that custom chips work best when they are tied to the rest of the cloud stack. Graviton plugs into managed services such as Amazon Aurora, Amazon Relational Database Service, and Amazon Elastic Kubernetes Service, while Trainium relies on the Neuron software stack and UltraClusters to scale model training across large server fleets. (aws.amazon.com) (aws.amazon.com) That leaves Amazon in a different position from a few years ago. Its cloud business is no longer only renting other companies’ chips; it is increasingly steering customers toward Amazon-designed silicon for both everyday infrastructure and the newest artificial intelligence workloads. (aboutamazon.sg) (aboutamazon.com)
